I am intending on installing two A100s on a dual Xeon CPU motherboard and utilizing NVlink. Do both A100s needs to be connected to the PCI-e lanes of the same CPU? Would the Nvlink functions still work, if both A100s are connected to separate CPUs?
NVIDIA recommends that datacenter GPUs be installed and configured by OEMs. That is the only supported configuration. Depending on your server, there is no guarantee that NVLink bridges will even fit in the server.
It is generally possible for NVLink to work even when GPUs are connected to different PCIE fabrics. DGX (-1, -2, -A100) servers have this configuration, for example.
There is no guarantee that anything will work in a server you configure yourself.
1 Like